How to Cut Token Costs with the Cursor Compass Router

Cursor’s upgraded request-level router allows you to dynamically assign language models based on task complexity, reducing inference costs by up to 68%. The official routing documentation details how the underlying Compass classifier evaluates prompts before execution. Here is how to configure the routing modes, leverage cache awareness, and balance cost against reasoning capabilities for your development workflow.

Understanding the Compass Classifier

Cursor Router abandons simple retry mechanisms and static fallbacks in favor of a request-level classifier. Before executing any generation, the router intercepts the request and passes it to a proprietary system called Compass. This system predicts the exact difficulty of the incoming change.

Compass evaluates four distinct input signals for every query. It analyzes the query itself to determine the language and intent of the prompt. It ingests the current context, encompassing the active file state and the broader codebase index. It predicts task complexity to gauge the required reasoning depth. Finally, it assesses the domain, distinguishing between specific programming languages or architectural layers like UI components versus backend services.

The system derives its routing logic from a training dataset of over 600,000 live developer requests. Compass optimizes for Average Follow-through Rate (AFC). This reward signal trains the system to favor models that generate code developers actually keep in their codebase, rather than optimizing for raw speed or generic benchmark scores. When evaluating AI output, AFC provides a more accurate reflection of utility than static tests.

Configuring Auto Routing Modes

You control the router by selecting one of three primary “Auto” modes in the Cursor model picker. Each mode targets a different baseline of performance and efficiency.

Mode	Priority	Target Models
Intelligence	Frontier quality for hard tasks	Claude Opus 5, Fable 5, GPT-5.6 Sol
Balance	High accuracy and low latency	Optimized daily-driver frontier models
Cost	Token efficiency	Grok 4.5, GPT-5.6 Luna Medium

The Auto Intelligence mode aggressively targets complex problems. As of the August 2026 update, this mode delivers user satisfaction scores higher than Fable 5 while operating at a 68% lower cost.

The Auto Balance mode serves as the recommended default for standard development. It outperforms Claude Opus 4.8 at a 41% lower cost.

The Auto Cost mode routes traffic to highly token-efficient models. Cursor explicitly utilizes Grok 4.5 as the primary price-efficient option in this tier. This mode helps reduce LLM API costs in production for straightforward boilerplate generation and repetitive refactoring.

Managing Cache Awareness

The Cursor Router operates as a cache-aware system. When evaluating a model switch mid-conversation, the router calculates the financial and latency overhead of discarding a warm prompt cache.

If the cumulative cost of a cache miss exceeds a predefined overhead cap, the router vetoes the model switch. This prevents the system from triggering expensive context reloads for minor queries that a cheaper model could handle, or conversely, switching to a cheap model when the cached context on a frontier model is already paid for. Understanding these context windows ensures the router maximizes your existing compute investments.

Task-Specific Routing Execution

The router excels at matching domain-specific tasks to specialized models. Benchmarks running on CursorBench 3.2 demonstrate strict behavioral patterns based on task categorization.

When a developer requests UI updates, the router automatically directs the prompt to models evaluated for high visual taste. Conversely, complex, long-horizon problems trigger routing to high-reasoning frontier models like GPT-5.6 Sol or Opus 5. This automated delegation eliminates the need for developers to manually swap models when transitioning from CSS adjustments to database architecture.

Enterprise early-access teams report total token savings of 30 to 50 percent across their development environments. You cannot manually adjust the Compass classifier weights, and explicitly selecting a static model in the UI bypasses the router entirely.

Set your workspace to Auto Balance to establish a baseline, then shift to Auto Intelligence when tackling architectural refactors that require maximum reasoning depth.